◆ Always-on Specialist
Quotes both sides of the book in microseconds, skews to its inventory, and widens or pulls as volatility moves. It works the flow inside the position limit the mandate sets; it does not size that limit. The fast model scores in-flight; the rationale is logged async to the surveillance control plane.
Memory
Working Live order book, current inventory, position vs. limit.
Episodic Today's fills and the name's intraday volatility regime.
Semantic Microstructure of each venue; tick sizes; auction mechanics.
Store Feature store + in-memory order-book state
Orchestration
MCP
Harness · Managed Agents … ultra-low-latency loop; pre-trade limit checks inline; async note-taking to an immutable surveillance log.
Tools
{ } Exchange/venue gateways (FIX) API { } Market-data feed (L2/L3) API { } Pre-trade risk-limit service API { } Surveillance event log API
Evals & guardrails
- Hard guardrail: cannot quote past position/loss limits … auto-flatten and halt on breach.
- Continuous spoofing/layering self-surveillance; anomalous patterns freeze the agent.
- Champion/challenger on every quoting-model update; shadow-quoted before promotion.
- Full OpenTelemetry trace of every quote + fill retained for MiFID II / Reg SHO audit.
Frontier edge
- ▲On-device / edge inference: the quoting model runs on co-located hardware so the decision-to-quote loop never leaves the latency budget.
- ▲Formal action-gating: every quote is cryptographically signed and replayable, so it provably cannot exceed the morning's risk limits.
- ▲Eval-gated continual learning (SEAL-style): absorbs each session's microstructure shifts offline, promoted only after shadow-quoting clears.
In numbers
120,000
Quotes / sec
1,800
Names quoted
+11 bps
Capture vs. spread
Handoffs
Across ⇢ Risk → Market Risk for intraday limit oversight